On Wed, Aug 14, 2024 at 10:20:55PM +1000, Michael Ellerman wrote: > Niklas Cassel <cassel@xxxxxxxxxx> writes: > > On Tue, Aug 13, 2024 at 10:32:36PM +1000, Michael Ellerman wrote: > >> Niklas Cassel <cassel@xxxxxxxxxx> writes: > >> > On Tue, Aug 13, 2024 at 07:49:34AM +0200, Jonáš Vidra wrote: > ... > >> >> ------------[ cut here ]------------ > >> >> kernel BUG at drivers/ata/pata_macio.c:544! > >> > > >> > https://github.com/torvalds/linux/blob/v6.11-rc3/drivers/ata/pata_macio.c#L544 > >> > > >> > It seems that the > >> > while (sg_len) loop does not play nice with the new .max_segment_size. > >> > >> Right, but only for 4KB kernels for some reason. Is there some limit > >> elsewhere that prevents the bug tripping on 64KB kernels, or is it just > >> luck that no one has hit it? > > > > Have your tried running fio (flexible I/O tester), with reads with a very > > large block sizes? > > > > I would be surprised if it isn't possible to trigger the same bug with > > 64K page size. > > > > max segment size = 64K > > MAX_DCMDS = 256 > > 256 * 64K = 16 MiB > > What happens if you run fio with a 16 MiB blocksize? > > > > Something like: > > $ sudo fio --name=test --filename=/dev/sdX --direct=1 --runtime=60 --ioengine=io_uring --rw=read --iodepth=4 --bs=16M > > Nothing interesting happens, fio succeeds. > > The largest request that comes into pata_macio_qc_prep() is 1280KB, > which results in 40 DMA list entries. > > I tried with a larger block size but it doesn't change anything. I guess > there's some limit somewhere else in the stack? > > That was testing on qemu, but I don't think it should matter? > > I guess there's no way to run the fio test against a file, ie. without a > raw partition? My real G5 doesn't have any spare disks/partitions in it. You can definitely run fio against a file. e.g. $ dd if=/dev/random of=/tmp/my_file bs=1M count=1024 $ sudo fio --name=test --filename=/tmp/my_file --direct=1 --runtime=60 --ioengine=io_uring --rw=read --iodepth=4 --bs=16M Perhaps try with 32M block size, so that it is larger than max segment size = 64K MAX_DCMDS = 256 256 * 64K = 16 MiB Perhaps also try with and without --direct. It could be interesting to use the page cache if you do --rw=readwrite that might possibly result in larger bios. Kind regards, Niklas